The Egyptian producer opens WhatsApp at 4:32 AM and dictates 40 seconds of Masry: what she needs, where, when, and what it costs. In the old world that voice note lives in her chat history and dies there.

In Tarkeeba, the voice note is the brief. Our pipeline runs it through Whisper (fine-tuned on Masry + Khaleeji), extracts entities into a structured schema, and turns the schema into a call sheet + budget + crew shortlist. Total time from tap-to-send: under 30 seconds.

The interesting design problem isn't transcription accuracy. It's the trust handoff: the producer has to see her brief and recognize it. That's why every AI-composed brief is editable, and every field shows the raw voice-note snippet it came from. Structure without erasure.